What changed
NVIDIA Nemotron 3.5 Super VL is now available in Amazon SageMaker JumpStart. The model can be deployed with a few steps in SageMaker Studio or with a few lines of the SageMaker Python SDK. It is a vision-language model (VLM) in the NVIDIA Nemotron 3.5 family, accepting images, video, and text as input and generating text. The model pairs a 120B-parameter hybrid Mixture-of-Experts (MoE) language backbone with 12B active parameters and a C-RADIO vision encoder. It supports context lengths of up to 1 million tokens and includes a built-in Multi-Token Prediction (MTP) head that supports speculative decoding. NVIDIA built Super VL for document intelligence, GUI understanding for computer-use agents, video question answering, and multimodal agentic applications.
What this means in practice
JumpStart offers two variants: NVFP4, quantized from BF16 with quantization-aware distillation for optimized inference, and the BF16 full-precision reference weights. The NVFP4 variant runs on G7 instances with a smaller GPU footprint, while the BF16 variant runs on P5en instances with NVIDIA H200 GPUs. The JumpStart default instance for NVFP4 is ml.g7.24xlarge, which serves as the recommended starting point for the generate profile. The model supports inputs such as image, video, and text, with outputs in text. It is intended for agentic workloads that must see and read, including vision-language agents that combine reasoning, coding, tool use, document understanding, and video understanding. Document intelligence includes OCR, charts, tables, and multipage documents. GUI understanding involves screen grounding and reasoning for computer-use agents that navigate applications. Video understanding allows analysis and summarization of short MP4 video clips, including answering questions about visual content. Multimodal agents support tool calling, JSON schema output through response_format, and configurable reasoning.
Limitations and availability
The JumpStart serving configuration does not enable speculative decoding, and the endpoint serves the base weights only, without loading LoRA adapters. The NVFP4 variant runs on ml.g7.24xlarge (default) and ml.g7.48xlarge, while the BF16 variant runs on ml.p5en.48xlarge. The checkpoint includes the MTP head, but the JumpStart serving configuration does not enable speculative decoding. The model is available under the OpenMDW License Agreement, version 1.1. The model-quality benchmarks report results for both checkpoints across reasoning, coding, agentic, long-context, and multimodal benchmarks, with evaluation settings held constant so that only the checkpoint differs. The NVFP4 variant matches BF16 on MMLongBench-Doc and stays within about one point on the others, scoring slightly higher on OCRBench v2 and OSWorld-Verified.
